Word Level Sign Language Recognition via Handcrafted Features

The ability to be understood and convey feelings, requests or ideas through words (spoken or written) is one of the most undervalue by all the humans who have the privilege to do it. Deaf community faces this challenge every single day and, even though, sign languages exist as way to battle against this issue, not all in deaf community knows who to use them; in fact, hearing community knows in a smaller proportion how to interpret them. By this reason sign language recognition area becomes relevant as an effort to solve this issue and create new communication channels. This work aims a methodology for word level sign language recognition, as principal highlights a small set of handcrafted features are defined, between them non-manual features are explored deeply. Data augmentation and dimensionality reduction were performed to obtain a concise feature space. Two recognition models were used (Bidirection Long Term Memory and Transformer) in LIBRAS dataset, and the best result was an accuracy of 94.33%, which was obtained with the bidirectional long term memory network.